TLDR: A new Social Welfare Function (SWF) Benchmark evaluates how large language models (LLMs) distribute scarce resources, balancing efficiency and fairness. The research reveals that general conversational ability is a poor predictor of an LLM’s allocation skill, most LLMs have a strong utilitarian bias prioritizing efficiency over equality, and their allocation strategies are highly vulnerable to external influences like output length constraints and social persuasion. This highlights the critical need for specialized benchmarks and targeted ethical alignment for LLMs in societal decision-making roles.
Large language models (LLMs) are increasingly being given significant responsibilities, including making high-stakes decisions that directly impact human welfare. However, the underlying principles and values that guide these powerful models when they distribute limited societal resources have largely remained a mystery.
To shed light on this crucial area, researchers have introduced the Social Welfare Function (SWF) Benchmark. This innovative simulation environment positions an LLM as a central allocator, tasked with distributing tasks among a diverse group of recipients. The benchmark is specifically designed to create a continuous challenge: maximizing collective efficiency (measured by Return on Investment or ROI) versus ensuring fair distribution (measured by the Gini coefficient).
The study evaluated 20 leading LLMs, creating the first-ever leaderboard for social welfare allocation. The findings offer three critical insights:
General Ability Doesn’t Equal Allocation Skill
Surprisingly, a model’s general conversational ability, as reflected in popular leaderboards, is a poor indicator of its skill in allocating resources. For instance, DeepSeek-V3-0324, which ranked 25th on a general Arena leaderboard, soared to 1st place on the SWF leaderboard. Conversely, models like Claude-4.1-Opus and GPT-5-High, top performers in general chat, ranked much lower on the SWF benchmark. This suggests that balancing fairness and efficiency requires a distinct capability not fully captured by standard conversational tests.
A Utilitarian Bias
Most LLMs exhibit a strong default tendency towards utilitarianism. This means they often prioritize maximizing overall group productivity and efficiency, even if it leads to significant inequality among recipients. They tend to focus on getting the most work done, sometimes at the expense of ensuring everyone gets a fair share of opportunities.
Also Read:
- Unpacking AI’s Social Compass: How Language Models Navigate Role Conflicts and Reveal Hidden Biases
- The Double-Edged Sword: How LLM Training Boosts Performance But Fosters Greed in Decision-Making
Vulnerability to External Influences
The allocation strategies of LLMs are highly susceptible to external factors. Simple changes, such as limiting the length of their output (forcing them to “think less”), can make them even more utilitarian, prioritizing efficiency over fairness. Additionally, social influence framing—like offering rewards for fairness (temptation) or imposing penalties for inequality (threats)—can significantly alter their allocation preferences. Direct incentives and disincentives proved to be the most effective in shifting LLM behavior towards fairer outcomes, more so than abstract appeals to normative standards.
The SWF score, which combines fairness and efficiency, highlights models that achieve a balanced approach. Top-performing models on this leaderboard excel by coordinating both objectives rather than maximizing just one. For example, DeepSeek-V3-0324 achieved its top rank by demonstrating a strong balance between fairness and efficiency.
These findings carry significant implications for the future of AI governance. They underscore the risks of deploying current LLMs in roles that involve critical societal decision-making without specialized evaluation and targeted ethical alignment. The research emphasizes the urgent need for benchmarks like the SWF to properly assess and guide AI systems that will increasingly shape human welfare. For more details, you can read the full paper here.
